Retail Credit Risk Model Validation: Performance and Stability Aspects

Abstract

Like a model without sufficient validation may only be a hypothesis (Stein 2002), a hypothesis without reason may just be an intuition. With the growing competition for acquisition of new and retention of existing retail banking customers, all stakeholders (financial regulators, auditors and business leads) have shifted their focus on accurate and timely validation of risk models. Intuitions and hypotheses are formulated into models at the time of development using statistical techniques, which get implemented into production system only after rigorous out-of-sample and out-of-time validation. Real test of the model, however, begins after implementation. In this paper we attempt to cover various aspects of validation of credit risk models after they get implemented. This paper is not an attempt to provide a laundry list of metrics but a conscious effort to add value to existing literature by focusing on application of appropriate methodologies and computation of relevant metrics for reviewers to comment on aspects of model performance and stability. Our approach looks at different dimensions of a credit risk model for detailed assessment. For such comprehensive validation we consider a) Power Testing b) Calibration and c) Sampling strategies All along the paper, we provide the readers with key validation aspects and corresponding metrics which serve as critical inputs for a model evaluator to approve or decline usage of implemented scores. Towards the end, the paper presents few illustrative snapshots of a recommended tool useful for analyzing model performance and stability results on regular basis.